I’ve ported ROCm 6.4.4 to gfx803 fully a couple months ago, then started patching 7.14 because I was targeting ONNX Runtime for smaller local models for Audiomuse AI. All that works, but I wanted to do some more broad testing and ended up trying llama.cpp with HIP (instead of Vulkan), as well as porting vLLM (from the gfx906 fork) to gfx803.
Now all of that is working and I just managed to stabilize ROCm 10.0, vLLM shipped in the repo still works against that. People have reported training a chess bot model on an RX 580 and image generation does seem to work more or less too.
I’m still looking into generating a more thorough test suite to make sure all rocBLAS/HIP kernels are working properly, but don’t expect me to submit anything upstream. Most fixes are pretty hacky and I have very little faith in getting these accepted.
So if you have an old mining rig and some 8-16 Polaris cards lying around, why not give it a try? vLLM is validated against Qwen 3.5-2B (only thing fitting on a single 8GB card that shares 3.8 architecture), so you could potentially run 27B or Flash Next on these cards.
And despite what the repo may say (Claude co-contributor to 2 commits), most of the AI assistance going into is has come from Deepseek V4 Flash on my local 170HX rig as well as some GLM 5.2